There are many circumstances in which there is a need to detect, identify, localise and track one or more non-cooperative objects of interest in some specified surveillance area. Such tasks can be performed by suitable active or passive sensors which can extract useful information by collaborative processing of signals reflected or emitted by those objects.
In contrast to applications employing active sensors, such as radar or active sonar, in which the surveillance region of interest is illuminated by an interrogating energy waveform to obtain object-backscattered returns, passive sensors capture only object-generated signals (or object-influenced signals from separate sources). For example, the movement of people, wheeled or tracked vehicles, speedboats or vibrating machinery can all generate wideband acoustic signals, which can be exploited for object detection, localisation and tracking.
As will be described in more detail below, an example in which object detection and localisation is useful is that of security surveillance with a network of distributed acoustic sensors forming an ‘acoustic fence’. When an object of interest, such as vehicle, has been detected and localised, the estimated object position can be utilized by security cameras for aiming and zooming in order to enhance the quality of recorded images. Such systems may be installed for monitoring purposes in industrial environments, e.g. to track moving obstacles, or to offer improved continuous surveillance of critical infrastructure, including power grids, power plants, gas and oil pipelines and water systems.
Another application is that of coastguard or littoral surveillance in which speedboats and other surface vessels of interest can be detected and localised by a network of floating buoys employing acoustic sensors and low-power radio transceivers providing an intersensor communication link.
In addition to the above surveillance and reconnaissance applications, in multimedia applications distributed microphone networks are capable of enhancing audio signals for improved intelligibility, and cuing for camera aiming.
When the distance between an acoustic source and the sensors is large, the direction of wave propagation is approximately equal at each sensor (the far-field condition), and the propagating field within the sensor network consists of planar waves. For a far-field source, only the direction-of-arrival (DOA) in the coordinate system of the sensors can be estimated directly from the relative delays of signals captured by sensors at different locations. Such relative delay is commonly referred to as the time difference of arrival, or simply TDOA.
The direction-of-arrival (DOA) measurement restricts the location of the source along a line in the estimated DOA. When multiple DOA measurements are made simultaneously by multiple spatially-separated sensors, a triangulation method may be used to determine a location of the source at the intersection of these lines-of-bearing.
When an acoustic source is close to the sensors (the near-field condition), the wave-front of the received signal is perceptibly curved with respect to the spatial extent of the sensor network. In this case, the propagation direction vectors at each sensor emanate from a common source location, and the set of all TDOA measurements can be utilized for localisation of the near-field source. Suitable algorithms utilizing TDOA measurements to determine the location of the near-field source are known to those skilled in the art.
Irrespective of the far/near-field source condition, the source location is always determined from TDOA estimates which need to be obtained from wideband acoustic signals captured by the sensors.
In a distributed sensor network, any two sensors will capture attenuated and time-shifted replicas, x1(t) and x2(t), of the same object-generated signal s(t), wherex1(t)=A1s(t)+n1(t)x2(t)=A2s(t−Δt)+n2(t)where A1 and A2 scale the amplitude of each signal, and Δt denotes the TDOA; waveforms n1(t) and n2(t) represent background noise and other interference.
The value of time difference of arrival Δt is usually determined by cross-correlating the wideband signals x1(t) and x2(t) captured by the sensors, i.e. by performing the operation
            R      12        ⁡          (      τ      )        =            1      T        ⁢                  ∫        0        T            ⁢                                    x            1                    ⁡                      (                          t              -              τ                        )                          ⁢                              x            2                    ⁡                      (            t            )                          ⁢                                  ⁢                  ⅆ          t                    where the integral is evaluated over the observation interval of duration T and for a range, −|Δtmax|<τ<|Δtmax|, of TDOA values of interest. The value of argument τ that maximises the cross-correlation function R12(τ) provides an estimate of an unknown TDOA.
In practice, prior to cross-correlation, the received signals may be suitably pre-filtered to accentuate frequencies for which signal-to-noise ratio (SNR) is highest and to attenuate background noise, thus increasing the resulting overall SNR. A cross-correlator utilizing signal pre-filtering is known in the prior art as a generalized cross-correlator.
The cross-correlation process, including pre-filtering, can also be implemented digitally, if sufficient sampling and quantising of the signal is used.
FIG. 1 is a block diagram of a known system cross-correlating a signal and its time-delayed replica to determine the value of TDOA. Each of the signals x1(t) and x2(t) is delivered to a respective filter 102, 104. The filtered version of signal x1(t) is passed through a variable delay line 106 to one input of a multiplier 108, the other input of which receives the filtered version of signal x2(t). The output of the multiplier 108 is integrated in a finite-time integrator 110. A peak detector 112 detects a peak in the output of the integrator 110. The position of this peak represents the delay time between the two signals x1(t) and x2(t).
Object-generated acoustic signals are classified as wideband signals since the ratio of their highest frequency component to lowest frequency component is relatively large. For example, for the audio range, 30 Hz to 15 kHz, the ratio is 500. In a case of wheeled and tracked vehicles, dominant frequency components may range from about 20 Hz to 2 kHz, resulting in a ratio of 100.
Not only do acoustic signals emitted by objects of interest occupy a wide frequency range, but they also will manifest a non-stationary and chaotic nature with identifiable intermittent transients. As a result, many known cross-correlation techniques based, explicitly or implicitly, on the assumptions of signal stationarity and noise Gaussianity are only of limited practical use. Furthermore, most practical implementations have to deal with discrete-time samples, so that the optimisation procedures and performance analyses carried out in the continuous-time framework cannot be fully applicable.
It would therefore be desirable to provide a method and an apparatus for determining time difference of arrival (TDOA) in a more efficient way than that provided by the prior art techniques.